When I compiled an onnx model, the error CUDA_ERROR_INVALID_PTX raised. I tried to change the target to target = tvm.target.cuda(model="nx")
and set_cuda_target_arch('sm_72')
, but it did not work.
It is possible to happen if you use too much shared memory. You might need autotuning before trying this out.
@junrushao I changed the target to target = tvm.target.create('llvm')
or target = tvm.target.cuda(model='tx2')
, the error did not raise, so It seems not causing by using too much shared memory.